JPH03141398A - Syntax processing device for continuous sound recognition - Google Patents

Abstract

PURPOSE:To decrease an enormous processing quantity without discarding a correct hypothesis by once storing an input word candidate of low probability and making syntax decision by returning to the stored word candidate when the probability that this word candidate can be connected to a word array candidate is low even if an acoustic distance value is small. CONSTITUTION:The probability that the input word candidate 11 is included in the result of voice recognition is first judged by a threshold deciding means 1 and the input word candidate 11 of the low probability is once stored in a word candidate memory means 8. The input word candidate 11 decided to be not connectable to the word array candidate 22 by a syntax deciding means 2 is connected before the word candidates 55 by a backward syntax deciding means 3 and further, whether this word candidate 55 can be connected behind the word array candidate 22 or not is decided by applying a backward consultation syntax rule 44. The input word candidate 11 which can be connected is additionally stored as the fresh word array candidate 22 to be connected to this word candidate 55 and the word array candidate 22. The processing quantity is decreased without lowering a sentence recognition rate in this way and the processing in real time is possible.

Description

[Detailed Description of the Invention] [Field of Industrial Application] This invention is a continuous speech recognition device that receives input word candidates from a sound processing device that acoustically processes input speech, and obtains speech recognition results by applying syntactic rules. Relating to a syntax processing device.

[Conventional technology]

Figure 7 shows, for example, literature (Kyoritsu Shuppan Information Science Course E, 1
9.3, Speech Recognition, pp. 139-142, 1979)
FIG. 2 is a functional block diagram of a conventional syntax processing device for continuous speech recognition shown in FIG.

In the figure, ■ indicates an input word candidate (01) obtained by acoustically processing input speech for a word string candidate (22) read out from a word string candidate storage means (50), and a syntax rule storage means (
A new word string candidate (22) is created by connecting the input word candidate 01) to the word string candidate (22) that satisfies the connection condition by applying the syntax rule (33) searched from 60) and performing a syntax test.
(40) is a syntax determining means that causes the word string candidate storage means (50) to additionally store the word string candidate storage means (50); recognition result determination means for reading word string candidates (22) and outputting them as speech recognition results (77);

(50) is a word string candidate storage means for storing a word string candidate (22) generated by syntactic processing, and (60) is a syntactic rule storage means for storing a syntactic rule (33) described in a syntactic state transition network. .

The above conventional syntax processing device for continuous speech recognition is as follows.

First, the syntax determining means (20) processes the input speech from the beginning along the time axis as shown in FIG.
Input word candidates (+1>
The syntax rule (33) retrieved from the syntax rule storage means (60) determines whether the word name W is accepted in the final state p of the word string candidate (22) read from the word string candidate storage means (50).
) to perform a syntactic test. For all of the word string candidates (sentence hypotheses) accepted and generated at this time, as shown in Figure 9, the words that constitute the words have a large total acoustic distance value (comprehensive evaluation value of acoustic distance values), that is, have a low likelihood. t! is adopted as the final word string candidate at that point and the next input word candidate 0
1) Beam search method (beam 5ear
ch method) to test the hypothesis.

In Figure 9, the numbers attached to the nodes indicate the likelihood, the circles indicate the hypotheses to be tested, and the black dots indicate the hypotheses to be rejected. As a result of this syntax test, a new word string candidate (22) formed by connecting the input word candidate (11) to a word string candidate (22) that satisfies the connection condition is additionally stored in the word string candidate storage means (50). do.

Next, when the utterance is finished and all the input word candidates (11) have been input, the recognition result determination means (40) selects the word string candidate (22) with the minimum total acoustic distance value from the word string candidate storage means (50).
) is read out and output as a voice recognition result (77).

In addition, for the method that uses the above-mentioned forward method as the processing direction, as shown in Fig. 10, first, all input word candidates (11) are memorized and the most probable word candidate is set as an initial hypothesis. The high-drive method (is l and driven me) in which the remaining parts are processed in the forward and backward direction of the time axis while extending from the beginning to the beginning or end of the sentence.
There is also a method using thocl).

[Problem to be solved by the invention]

In the conventional syntax processing device for continuous speech recognition as described above,
Since the forward-looking method processes hypothesis evaluation and verification in the time axis direction, if a correct hypothesis is rejected midway through, correct recognition results cannot be obtained, and word candidates near the beginning of the utterance have acoustic distance values. is not necessarily small, and it is necessary to formulate many hypotheses in order to obtain correct recognition results.In order to further reduce rejection and improve sentence recognition rate, it is necessary to formulate many hypotheses even for word candidates with relatively large acoustic distance values. The amount of processing required will be enormous.

On the other hand, the method using the island driving method is less likely to reject the correct hypothesis at the beginning of processing and can be expected to have a high recognition rate. Since all the syntactic states to be used are hypotheses, the amount of processing required to verify them becomes large. In addition, real-time processing is not possible because the initial hypothesis must be determined after all word candidates in the uttered sentence are detected.

Conventional technology has had the problems described above.

The problem to be solved by the present invention is to provide a continuous speech recognition device using the syntax processing device of the present invention that reduces the processing amount without reducing the sentence recognition rate compared to the conventional one.

and to enable real-time processing.

[Means to solve the problem]

In order to achieve the above object, a syntax processing device for continuous speech recognition according to the present invention is characterized in that it includes the following means.

A continuous speech recognition device acoustically processes input speech. In the syntax processing device, which receives input word candidates from the acoustic processing device and applies syntactic rules to obtain speech recognition results, the word string candidates generated by hq syntactic processing are stored. A word string candidate storage means, and a syntax rule storage means for storing syntactic rules described in a syntactic state transition network represented by a certain state, a word name that can be accepted from that state, and a state to which the word transitions when the word is accepted.

a reverse syntax rule storage means for storing a reverse syntax rule described in a syntactic state transition network represented by a certain word name, a word name connected before the word, and a state that transitions when connected; a word candidate storage means for storing word candidates that are unlikely to belong to a word string candidate, and once the input word candidate is subjected to syntax determination by determining a threshold value of the acoustic distance value of the input word candidate, the word candidate storage means is provided. a threshold value determining means for determining whether the input word candidate from the threshold value determining means can be connected to the word string candidate storage means; a syntax retrieved from the syntax rule storage means; Syntax determining means for determining syntax by applying rules and determining whether to additionally store the word string candidate as a new word string candidate in the word string candidate storage means or to further perform backward syntax determination on the input word candidate; The word candidates to which the input word candidates from . backward syntax determination means for reading from the column candidate storage means, performing backward syntax determination by similarly applying reverse lookup syntax rules, and determining whether or not to cause the word string candidate storage means to additionally store a new word string candidate; When the acoustic processing of the input speech is completed and an acoustic processing end signal is input, a word string candidate whose terminal position is near the terminal position of the input speech and whose total acoustic distance value is the minimum is read out from the word string candidate storage means and output as a speech recognition result. A recognition result determination means is provided.

[For production]

The syntax processing device for continuous speech recognition configured as above is
First, the probability that an input word candidate obtained by acoustically processing input speech is included in the speech recognition result is determined by a threshold value determination means, and input word candidates with a low probability are temporarily stored.

Next, the syntax determining means determines whether the input word candidate determined by the threshold determining means as having a high possibility of being included in the speech recognition result can be connected after the word string candidate by applying a syntactic rule, Connectable input word candidates are additionally stored as new word string candidates connected to the word string candidates.

Further, the backward syntax determining means stores input word candidates that are determined by the syntax determining means to be unconnectable to the word string candidate, and connects them before the word candidate, and furthermore, the word candidates can be connected after the word string candidate. It is determined whether or not this is the case by applying reverse lookup syntax rules, and connectable input word candidates are additionally stored as new word string candidates formed by connecting the word candidates and the word string candidates.

Finally, when the acoustic processing end signal is input to the recognition result judgment means, the terminal position of the t word string candidates generated by the above syntax processing is near the terminal position of the input speech and the total acoustic distance value is the minimum. Output as result.

〔Example〕

FIG. 1 is a functional block diagram of a syntax processing device for continuous speech recognition showing one embodiment of the present invention.

In the figure, (1) determines the threshold value of the acoustic distance value of the input word candidate (11) obtained by acoustically processing the input speech,
Threshold determination means determines whether the input word candidate (II) is subjected to syntactical judgment or whether it is temporarily stored as a word candidate. Adjacent to the end position of the word string candidate (22) read from the string candidate storage means (5), further (
The syntax rule (3) indicates that the word name W of the input word candidate 01) is accepted in the final state p of the word string candidate (22).
3) can be retrieved from the syntax rule storage means (6), the threshold value of the total acoustic distance value of a new word string candidate formed by connecting the word string candidate (22) and the input word candidate 01) is determined, and a new The syntax+U-determining means (3) determines whether to additionally store the word string candidate (22) as a word string candidate (22), or to further perform backward syntax judgment on the input word candidate 01) since there is no connectable word string candidate. The starting position of the input word candidate (11) is adjacent to the ending position of the word candidate (55) read from the word candidate storage means (8), and the starting position of the word candidate (55) is the word string candidate storage means. Adjacent to the end position of the word string candidate (22) read from (5), and also the word candidate? ! The reverse syntax rule (44) indicating that the word name W' of the word candidate (55) and the word name W of the subsequent input word candidate 01) are connected in the final state S of (22) is stored as a reverse syntax rule. When it is possible to search from means (7), the word string candidate (22) and word candidate (55)
and a backward syntax judgment that determines the threshold value of the total acoustic distance value of a new word string candidate formed by connecting the input word candidate (11) and determines whether or not to additionally store it as a new word string candidate (22). Means (4) indicates that the end position of the word string candidate generated by the syntactic processing is near the end position of the input speech when the acoustic processing end signal (66) is input after all input word candidates θ1) have been input. The one with the minimum total acoustic distance value is read out from the word string candidate storage means (5) and the speech recognition result (
77), and (5) outputs the word string candidate (22) generated by the syntactic processing as shown in FIG. 2, the address, word name, word acoustic distance value, Word string comprehensive acoustic distance value, starting position, and ending position U.

The word string candidate storage means stores the number of words from the beginning of the sentence, the state, and the address in which the previous word of each word is stored.(6) stores a certain state p, the word name W that can be accepted from that state, and the word The state q to which it transitions when it is accepted is the third
Syntax rules (33) described using a syntax state transition network as shown in the figure
The syntax rule storage means is configured to store in advance a certain state pl and a word name fllJ, and to search for a transition state qI7 by comparing it with the first two words using a certain state pl and a word name fllJ as keys, as shown in FIG.

(7) describes a certain word name W, the word 8' that can be connected before that word, and the state syP+q that it transitions to when connected, using a syntactic state transition network as shown in Figure 5, and then uses t reverse lookup syntactic rules. (44) in advance, and word name 1 as shown in Figure 6.
Before fl is the word name W' (4 can be connected and the resulting transition state is ff1Jll) lJII? A reverse syntax rule storage means configured to be able to search backwards in the time axis to find that lj, (8) stores word candidates (II) that are unlikely to belong to the word string candidate (22). This is a word candidate storage means.

The syntax processing device for continuous speech recognition according to the above embodiment is as follows.

First, an acoustic processing device that acoustically processes input speech using a continuous speech 4D device generates word names, starting and ending positions, and sound pressure g.
An input word candidate θl) including each information of l value is received.

Next, when the acoustic distance value of the input word candidate Of) is less than the threshold value, the threshold value determination means (1) determines that there is a high possibility that the input word candidate Of) is included in the speech recognition result (77), and the syntax determination means (2) ) is sent the input word candidate 01). On the other hand, when the value exceeds the value, the input word candidate (II> is temporarily stored in the word candidate storage means (8) as the possibility of similar occurrence is low. Also, the starting position of the input word candidate Of) is determined by the syntax determination means (2). is adjacent to the end position of the word string candidate (22) read from the word string candidate storage means (5), and the word name w1J of the input word candidate (11>) is accepted in the final state p+ of the word string candidate (22). When the syntactic rule (33) indicating that the syntax rule (33) that When the value is less than the threshold, it is determined that the input word candidate OD can be connected to the read word string candidate (22), and it is additionally stored in the word string candidate storage means (5) as a new word candidate?ll! ('22). On the other hand, if there is no word string candidate that can connect the input word candidate 01> beyond the interval value, the input word candidate (11) is sent to the backward syntax determining means (3).

Furthermore, input word candidates (11) are determined by backward syntax determination means (3).
The start position of the word candidate (55) is adjacent to the end position of the word candidate (55) read from the word candidate storage means (8), and the word candidate (
The starting position of the word string candidate (55) is adjacent to the end position of the word string candidate (22) read from the word string candidate storage means (5), and furthermore, the word candidate (55) is in the final state S, j of the word string candidate (22).
) is connected to the word name wl of the previous input word candidate (11).
4) can be retrieved from the reverse syntax rule storage means (7).

The word string candidate (22), word candidate (55), and input word candidate? If the total acoustic distance value of a new word string candidate formed by connecting lfl (+1>) is less than or equal to the threshold, input word candidate 01
) read out 17 word candidates (55) and word string candidates (22)
It is determined that the word string candidate can be connected to the word string candidate storage means (5), and is additionally stored as a new word string candidate (22).

The above threshold value judgment is made for all input word candidates 01).

After performing syntax determination and backward syntax determination processing and completing all inputs, a sound processing end signal (66) is received from the sound processing device.

Finally, the recognition result determination means (4) uses the word string candidate storage means (
From 5), the word string candidate (22) generated in the above process whose terminal position is near the terminal position of the input speech and whose overall acoustic distance value is the minimum is read out and output as a speech recognition result (77).

〔Effect of the invention〕

This invention is constructed as described above.

Instead of performing syntax judgment on all input word candidates as in the past, only input word candidates with a small acoustic distance value and a high probability of being included in the speech recognition result are subjected to syntax judgment, and those with low probability are temporarily memorized. Leave it behind.

Even if the acoustic distance value is small, when the possibility of connecting to a word string candidate is low, we go back to the memorized word candidate and make a syntax judgment, which eliminates the problem with the conventional method using the forward method. The huge amount of processing required can be reduced without rejecting correct hypotheses. In addition, although there is some backward processing, it is basically processed in the time axis direction, so processing cannot be started until after the utterance is finished, which was a problem with the conventional high drive method. This can also be resolved.

Therefore, the continuous speech recognition apparatus using the syntax processing apparatus of the present invention has the effect of reducing the amount of processing without lowering the sentence recognition rate and enabling real-time processing compared to conventional apparatuses.

[Brief explanation of the drawing]

FIG. 1 is a functional block diagram of a syntax processing device for continuous speech recognition showing one embodiment of the present invention, and FIG. 2 is a diagram showing an example of the storage configuration of the word string candidate storage means of FIG. 1. 3 and 4 are diagrams showing a syntactic state transition network describing syntactic rules in the syntactic rule storage means of FIG. 1 and an example of its storage configuration;
5 and 6 are diagrams showing a syntactic state i transition network that describes reverse syntax rules in the reverse syntax rule storage means of FIG. 1 and an example of its storage configuration, and FIG. FIGS. 8, 9, and 10, which are functional block diagrams of the syntax processing device for recognition, are diagrams showing a system using the target method, beam search method, and high drive method in the conventional example of FIG. 7. In the figure, (1) is a threshold determination means, (2) is a syntax determination means, (3) is a backward syntax determination means, (4) is a recognition result determination means, (5) is a word string candidate storage means, and (6) is a Syntax rule storage means, (7) is a reverse syntax rule storage means, (8)
is word candidate storage means, (+1> is input word candidate, (
22) is a word string candidate, (33) is a syntax rule, (44
1 is a reverse syntax rule. (55) is a word candidate, (66) is a sound processing end signal,
(77) Voice recognition result. Note that in the two figures, the same reference numerals indicate the same or equivalent parts.

Claims (1)

[Claims] A continuous speech recognition device acoustically processes input speech. Receives input word candidates from the acoustic processing device and applies syntactic rules to obtain speech recognition results. In the syntactic processing device, word string candidates generated by syntactic processing are stored. a word string candidate storage means; a syntactic rule storage means for storing syntactic rules described in a syntactic state transition network represented by a certain state, a word name that can be accepted from that state, and a state to which the word transitions when the word is accepted; , a reverse syntax rule storage means for storing a reverse syntax rule described in a syntactic state transition network represented by a word name, a word name that can be connected before the word, and a state to which it transitions when connected; and a word string. a word candidate storage means for storing word candidates that are unlikely to belong to the candidates, and a threshold value judgment of an acoustic distance value (difference between an acoustic feature amount of input speech and a reference value of a standard pattern) of the input word candidate is provided. a threshold value determining means for determining whether the input word candidate is subjected to syntax determination or to be temporarily stored in the word candidate storage means; Read the word string candidate from the word string candidate storage means, apply the syntax rules retrieved from the syntax rule storage means to determine the syntax, and either cause the word string candidate storage means to additionally store the word string candidate as a new word string candidate, or further store the input word candidate. a syntax determining means for determining whether to perform backward syntax determination; a word candidate to which the input word candidate from the syntax determining means can be connected is read from the word candidate storage means; and a reverse syntax rule retrieved from the reverse syntax rule storage means is read out. Then, a word string candidate to which the word candidate can be connected is read out from the word string candidate storage means, a backward syntax is determined by similarly applying the reverse syntax rules, and the word string candidate storage means is used as a new word string candidate. A backward syntax determining means determines whether or not to additionally store the input speech, and when an acoustic processing end signal is input after completing the acoustic processing of the input speech, the word string candidate storage means determines the end position of the input speech when the end position is near the end position of the input speech. 1. A syntax processing device for continuous speech recognition, comprising recognition result determination means for reading out a word string candidate with a minimum distance value and outputting it as a speech recognition result.